A cross-modal large model driven ceramic cultural product marketing system

Abstract

This invention discloses a cross-modal large-model driven marketing system for ceramic cultural products, relating to the field of online marketing. It includes: a data collection module for collecting and filtering cross-modal representation data of ceramic cultural products from online sales platforms, as well as full-volume interactive behavior data of corresponding users within the platform; and an encoding and mapping module for performing unified-dimensional cross-modal feature encoding and semantic mapping processing on the collected and filtered cross-modal data of ceramic cultural products and user interactive behavior data. This invention accurately mines users' personalized needs for ceramic cultural products and adapts them to product attributes. It combines a hierarchical matching strategy prioritizing ceramic cultural attributes to achieve targeted product recommendations. Furthermore, it can update matching results in real time based on user behavior and optimize product display according to push adaptation values, making online recommendations for ceramic cultural products more aligned with user needs and improving marketing accuracy and user experience.

Description

Technical Field

[0001] This invention relates to the field of online marketing technology, specifically to a cross-modal, large-scale model-driven marketing system for ceramic cultural products. Background Technology

[0002] Online product marketing is a product promotion and sales model based on internet channels. At present, ceramic cultural products and handicrafts are also gradually adopting this method of marketing to increase their exposure in the public's daily life.

[0003] Patent application number 202210907698.X discloses a marketing evaluation system for cultural and creative products. This application aims to address the problem that "current cultural and creative product trading platforms mainly consist of platforms for physical cultural and creative products, platforms for virtual products (digital products), and e-commerce platforms for marketing and promoting cultural and creative services. These platforms have too narrow a target audience and lack dedicated categories or common search features for products, resulting in scattered homepage categories and cumbersome procedures for sellers to find the cultural products they need. Furthermore, the display of cultural and creative products needs careful consideration, as these products involve comprehensive appearance and the cultural connotations of various elements."

[0004] For ceramic cultural products, due to their fragility, it is difficult to promote them in many places through offline marketing. Obviously, online marketing is more suitable for ceramic cultural products. However, when ceramic cultural products are listed on online sales platforms, how to generate ceramic cultural product recommendations that are suitable for users based on the limited information of users on the online sales platform is a problem that has not yet been addressed or solved. As a result, online sales platforms mainly push ceramic cultural products with popular or trending appeal, thus failing to improve the sales rate of ceramic cultural products.

[0005] To address this, we propose a cross-modal, large-scale model-driven marketing system for ceramic cultural products. Summary of the Invention

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a cross-modal large model driven marketing system for ceramic cultural products, which can effectively solve the problems of the existing technology.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions; This invention discloses a cross-modal large-model driven marketing system for ceramic cultural products, comprising: The system comprises the following modules: a data collection module for collecting and filtering cross-modal representation data of ceramic cultural products from online sales platforms, as well as full-volume user interaction behavior data; an encoding and mapping module for performing unified-dimensional cross-modal feature encoding and semantic mapping on the collected and filtered cross-modal data of ceramic cultural products and user interaction behavior data; a generation module for generating a set of personalized user demand tags corresponding to the consumption dimension of ceramic cultural products based on the encoded and mapped user interaction behavior feature data; a construction module for constructing a set of product attribute tags that matches the set of personalized user demand tags based on the encoded and mapped ceramic cultural product feature data; a matching module for generating a targeted recommendation list of ceramic cultural products for a single user based on the real-time matching relationship between the set of personalized user demand tags and the set of product attribute tags; and a push module for pushing the generated targeted recommendation list to the corresponding user's online sales platform interface. The acquisition module is interconnected with an encoding and mapping module via a wireless network. The encoding and mapping module is interconnected with a generation module and a construction module via a wireless network. The construction module is interconnected with a matching module via a wireless network. The matching module is interconnected with a push module via a wireless network. The push module is interconnected with the acquisition module via a wireless network.

[0008] Furthermore, the cross-modal representation data of ceramic cultural products collected by the acquisition module includes visual modal data, text modal data, and cultural semantic modal data specific to ceramic cultural products; The visual modal data includes image feature data of ceramic shapes, patterns, and glazes; the text modal data includes textual descriptions of ceramic kilns, techniques, and cultural connotations; and the cultural semantic modal data includes semantic feature data of the intangible cultural heritage school to which the ceramics belong and their historical and cultural background. The full user interaction behavior data includes time-series data of user behavior such as browsing, collecting, adding to cart, placing orders, commenting, and sharing ceramic culture products. The filtering operation of the collection module is to perform an initial screening of cross-modal representation data based on the relevance of ceramic culture attributes, and a secondary screening of user interaction behavior data based on the validity of behavior, removing product data that is not associated with ceramic culture attributes and user data that has no valid interaction behavior.

[0009] Furthermore, the encoding and mapping module performs cross-modal feature encoding and semantic mapping processing, including: Feature Encoding: Feature extraction and encoding were performed on the visual modal data, textual modal data, cultural semantic modal data, and time-series data of all user interaction behaviors of ceramic cultural products. Multi-layer convolutional feature extraction and dimensionality reduction were performed on the image feature data of ceramic shapes, patterns, and glazes to obtain visual modal feature vectors. The textual descriptions of ceramic kilns, techniques, and cultural connotations are segmented and encoded using cultural semantic terms to obtain text modal feature vectors. The semantic feature data of ceramic intangible cultural heritage schools and historical and cultural backgrounds are linked with cultural semantic entities and quantified and encoded to obtain cultural semantic modal feature vectors. The user behavior modal feature vector is obtained by performing temporal feature serialization encoding and behavior validity weight quantization on the time-series data of user browsing, favorites, adding to cart, placing orders, commenting, and sharing behaviors. ; Unified Dimension Feature Transformation and Fusion: A unified dimension feature space for ceramic culture product marketing is pre-constructed, and modal feature projection matrices for visual, textual, cultural semantics, and user behavior are configured respectively. , , , By projecting the corresponding projection matrix onto this unified-dimensional feature space, we obtain the same-dimensional feature vectors for each modality. Then, the attention weights contributing to the marketing of ceramic cultural products are calculated using a cross-modal attention mechanism. Based on these weights, the attention weights of product category feature vectors of the same dimension are then applied. The features are fused to generate a unified cross-modal fusion feature vector for the product. For user behavior feature vectors of the same dimension By optimizing the semantic relevance of product category features, a unified cross-modal fusion feature vector for the user end is generated. ; Semantic mapping: Constructing a set of anchor points for ceramic cultural attributes, which includes at least four core ceramic cultural attributes: intangible cultural heritage certification, kiln origin and school, craftsmanship, and cultural connotation. Using this anchor point set as a semantic reference benchmark, cross-modal fusion feature vectors from the product side are then used. Cross-modal fusion feature vectors with the user end Synchronously map to the preset semantic space of ceramic culture product marketing features, and then correct the coordinate distribution of feature vectors in the semantic space through the semantic similarity calculation of ceramic culture attributes, so as to complete the semantic alignment of product features and user features in the same semantic space.

[0010] Furthermore, when calculating the attention weights for the contribution of each modal feature to the marketing of ceramic cultural products through the cross-modal attention mechanism, the following applies: ; In the formula: The attention weights are contributed to the modal feature vector of the m-th product class in the same dimension. Let be the correlation degree of the ceramic culture attribute of the m-th modal feature; Let be the marketing feature expressibility of the m-th modality.

[0011] Furthermore, the user behavior feature vector of the same dimension The optimization operation based on the semantic relevance of product category features is as follows: calculate the semantic relevance of the feature values ​​of each dimension of the user behavior feature vector with the ceramic culture attribute of the cross-modal fusion feature vector of the product, and use this relevance as the dimension optimization coefficient to correct the feature values ​​of each dimension of the user behavior feature vector dimension by dimension. By calculating the semantic similarity of ceramic cultural attributes, when correcting the coordinate distribution of feature vectors in the semantic space, the semantic similarity between the cross-modal fusion feature vectors of the product end and the user end and the features of each anchor point in the ceramic cultural attribute feature anchor point set is calculated. Taking the semantic weight of each cultural attribute anchor point in the marketing of ceramic cultural products as the correction benchmark, the coordinate values ​​of each feature vector in the marketing-specific feature semantic space are shifted and corrected towards the coordinate direction of the cultural attribute anchor point with high semantic similarity. The correction process keeps the dimension and modulus of the feature vector unchanged.

[0012] Furthermore, the generation module generates a set of personalized user demand tags for the consumption dimension of ceramic cultural products based on the encoded and mapped user feature data: Select the consumption dimension of ceramic culture products, then calculate the matching degree between user feature data and the demand of each consumption dimension, and finally take the consumption dimension with the demand matching degree higher than the preset threshold as the core demand dimension to generate the corresponding personalized demand tag set. The consumption dimensions of ceramic cultural products include cultural collection, craftsmanship appreciation, daily use, gift-giving, and intangible cultural heritage collection. The degree of demand matching... , This represents the normalized user cross-modal fusion feature vector after encoding mapping. Let represent the standardized feature basis vector of the k-th consumption dimension of ceramic cultural products. This indicates the total number of cross-modal types, corresponding to the three modalities of visual, textual, and cultural semantics. Let represent the intermodal correlation coefficient between the m-th cross-modal feature and the k-th consumption dimension. This represents the number of dimensions in the unified feature space. This represents the feature space projection coefficient of the s-th dimension feature in the k-th consumption dimension.

[0013] Furthermore, the construction module constructs a product attribute tag set that matches the user's personalized demand tag set based on the encoded and mapped ceramic culture product feature data: Based on the cross-modal fusion feature data of the product after encoding and mapping, the features are clustered and dimensionalized according to the exclusive attribute dimensions of ceramic culture products to obtain multi-dimensional attribute feature clusters of the product. Then, according to the core requirement dimensions of the user personalized requirement tag set, the multi-dimensional attribute feature clusters of the product are adapted and filtered to remove attribute features that have no semantic relationship with the core requirement dimensions. Finally, the filtered attribute features are tagged to generate product attribute tag set. The exclusive attribute dimensions of the ceramic cultural products include kiln site attributes, craftsmanship attributes, vessel type attributes, cultural connotation attributes, material grade attributes, and intangible cultural heritage certification attributes.

[0014] Furthermore, when the matching module generates a targeted recommendation list of ceramic culture products for a single user, it performs the generation operation based on a hierarchical matching strategy that prioritizes ceramic culture attributes: The first level matches the core cultural attributes of ceramic culture and intangible cultural heritage certification; the second level matches the technological attributes of kiln and craftsmanship; and the third level matches the practical attributes of vessel shape and material. The matching module first calculates the semantic matching degree of the tags under each level, and then obtains the comprehensive matching degree between the product and the user based on the weight of the levels. At the same time, it triggers matching updates based on real-time changes in user interaction behavior. Finally, it sorts the comprehensive matching degree from high to low and generates a targeted recommendation list of ceramic culture products for each user.

[0015] Furthermore, when the push module pushes the targeted recommendation list to the user interface, it first calculates the push adaptation value for each ceramic culture product in the recommendation list, and then determines the display position, display section and display frequency of the product on the interface based on the push adaptation value. After the push module completes the push operation, it refreshes the system operation at a specified time interval. ; In the formula: The push adaptation value for the i-th ceramic culture product in the targeted recommendation list; Let be the feature vector representing the probability of reaching the target user for the i-th ceramic cultural product; Let be the feature vector of the fit between the i-th ceramic culture product and the user demand tag; Let i represent the cultural dissemination degree of the i-th ceramic cultural product; The cross-scenario push adaptation coefficient for the i-th ceramic culture product; This is the element-wise dot product operation of the eigenvectors.

[0016] Compared with the known prior art, the technical solution provided by this invention has the following beneficial effects: This invention, through targeted collection and two-layer screening of multimodal representation data and user interaction behavior data of ceramic cultural products, can accurately eliminate product data without ceramic cultural attributes and user data without effective interaction, thus consolidating the accuracy and effectiveness of data application. At the same time, by independently encoding multimodal features and fusing them with unified-dimensional feature transformation, and combining cross-modal attention mechanisms to assign appropriate marketing contribution weights to each modality feature, it enables precise semantic alignment of product and user features in the exclusive semantic space of ceramic cultural product marketing, allowing feature representations to fit the exclusive attributes and consumer demand dimensions of ceramic cultural products. This invention accurately mines users' personalized core needs in the consumption of ceramic cultural products based on feature data, constructs product attribute tags that are highly adapted to these needs, generates a recommendation list with a hierarchical matching strategy that prioritizes cultural attributes and combines it with real-time changes in user behavior, and dynamically adjusts the hierarchical weight according to the semantic matching degree to maintain the fit between product recommendations and user needs. This invention optimizes the display position, section, and frequency of products on the interactive interface by calculating and pushing adaptation values, so that product reach is more in line with users' browsing habits and effectively improves the online marketing conversion efficiency of ceramic cultural products. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0018] Figure 1 This is a schematic diagram of the structure of a cross-modal large-scale model-driven marketing system for ceramic cultural products. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] The present invention will be further described below with reference to embodiments.

[0021] Example: This embodiment presents a cross-modal, large-model-driven marketing system for ceramic cultural products, such as... Figure 1 As shown, it includes: The data collection module is used to collect and filter cross-modal representation data of ceramic cultural products on online sales platforms, as well as the full interactive behavior data of corresponding users within the platform. The cross-modal representation data of ceramic cultural products collected by the acquisition module includes visual modal data, text modal data, and cultural semantic modal data specific to ceramic cultural products; Visual modal data includes image feature data of ceramic shapes, patterns, and glazes; text modal data includes textual description data of ceramic kilns, techniques, and cultural connotations; and cultural semantic modal data includes semantic feature data of the intangible cultural heritage school to which the ceramics belong and their historical and cultural background. The full user interaction behavior data includes time-series data of user behavior such as browsing, collecting, adding to cart, placing orders, commenting, and sharing ceramic culture products. The filtering operation of the collection module is to perform an initial screening of cross-modal representation data based on the relevance of ceramic culture attributes, and a secondary screening of user interaction behavior data based on the validity of behavior, removing product data that is not related to ceramic culture attributes and user data that has no effective interaction behavior. During the initial screening, core feature anchors of ceramic cultural attributes are constructed based on ceramic intangible cultural heritage schools, kilns, historical and cultural backgrounds, and cultural connotations. Then, the semantic matching degree between the cross-modal semantic similarity data of each ceramic cultural product and the core feature anchors is calculated using a cross-modal semantic similarity algorithm. Product data with no association with ceramic cultural attributes that have a semantic matching degree lower than a preset threshold are removed. During the secondary screening, the validity of each user's interaction with ceramic culture products is determined. The validity of each interaction includes the duration of the interaction, the frequency of the interaction, and the depth of the interaction. The user behavior data of each dimension is then standardized and quantified, and the validity of a single type of user interaction and the validity of the user's overall interaction are calculated. User data with no valid interaction that is below the preset threshold is removed. The encoding and mapping module is used to perform unified-dimensional cross-modal feature encoding and semantic mapping processing on the collected and filtered cross-modal data of ceramic cultural products and user interaction behavior data. The encoding and mapping module performs cross-modal feature encoding and semantic mapping processing, including: Feature Encoding: Feature extraction and encoding were performed on the visual modal data, textual modal data, cultural semantic modal data, and time-series data of all user interaction behaviors of ceramic cultural products. Multi-layer convolutional feature extraction and dimensionality reduction were performed on the image feature data of ceramic shapes, patterns, and glazes to obtain visual modal feature vectors. The textual descriptions of ceramic kilns, techniques, and cultural connotations are segmented and encoded using cultural semantic terms to obtain text modal feature vectors. The semantic feature data of ceramic intangible cultural heritage schools and historical and cultural backgrounds are linked with cultural semantic entities and quantified and encoded to obtain cultural semantic modal feature vectors. The user behavior modal feature vector is obtained by performing temporal feature serialization encoding and behavior validity weight quantization on the time-series data of user browsing, favorites, adding to cart, placing orders, commenting, and sharing behaviors. ; Unified Dimension Feature Transformation and Fusion: A unified dimension feature space for ceramic culture product marketing is pre-constructed, and modal feature projection matrices for visual, textual, cultural semantics, and user behavior are configured respectively. , , , By projecting the corresponding projection matrix onto this unified-dimensional feature space, we obtain the same-dimensional feature vectors for each modality. Then, the attention weights contributing to the marketing of ceramic cultural products are calculated using a cross-modal attention mechanism. Based on these weights, the attention weights of product category feature vectors of the same dimension are then applied. The features are fused to generate a unified cross-modal fusion feature vector for the product. For user behavior feature vectors of the same dimension By optimizing the semantic relevance of product category features, a unified cross-modal fusion feature vector for the user end is generated. ; Semantic mapping: Construct a set of anchor points for ceramic cultural attributes. The anchor point set should include at least four core ceramic cultural attributes: intangible cultural heritage certification, kiln school, craftsmanship, and cultural connotation. Using the anchor point set as a semantic reference benchmark, cross-modal fusion feature vectors from the product side are then used. Cross-modal fusion feature vectors with the user end Synchronously map to the preset semantic space of ceramic culture product marketing features, and then correct the coordinate distribution of feature vectors in the semantic space through the semantic similarity calculation of ceramic culture attributes, so as to complete the semantic alignment of product features and user features in the same semantic space. Among them, the behavior effectiveness weight quantification is achieved by setting basic weights corresponding to the correlation between each interactive behavior such as browsing, favorites, adding to cart, placing orders, commenting, and sharing and the consumption decision of ceramic cultural products. Then, the actual execution frequency and duration of each behavior are normalized across the entire domain. The normalized frequency value and duration value of each behavior are then multiplied by the basic weight of the behavior in turn. The two product results are then added together to obtain the effective quantification weight value of each single interactive behavior. Each interactive behavior is calculated in this way to form a complete set of behavior effectiveness weights. When calculating the contribution weights of each modal feature to the marketing of ceramic cultural products using a cross-modal attention mechanism, the following rules apply: ; In the formula: The attention weights are contributed to the modal feature vector of the m-th product class in the same dimension. Let be the correlation degree of the ceramic culture attribute of the m-th modal feature; The marketing feature expressibility of the m-th modality; This formula uses the product of the correlation degree of ceramic culture attributes and the expression degree of marketing features of the same dimension modal features of each product category as the core contribution value of a single modal feature. Then, by summing and normalizing the core contribution values ​​of all modal features, the actual contribution attention weight of each modal feature is obtained. This is different from the equal allocation of weights in traditional cross-modal fusion. Starting from the dual core dimensions of correlation of ceramic culture attributes and expression of marketing features, the weight allocation can accurately match the marketing attributes of ceramic culture products, thereby truly reflecting the actual role and value of different modal features in the marketing of ceramic culture products. in, The feature proportion of the core attributes of ceramic culture in this modality is quantified. Based on the specific consumption dimension of ceramic cultural products, the feature coverage of this modal feature to each consumption dimension is calculated and the average value is obtained. The feature coverage is the ratio of the number of core features of the corresponding consumption dimension that can be matched in this modal feature to the total number of core features of this consumption dimension. User behavior feature vectors of the same dimension The optimization operation based on the semantic relevance of product category features is as follows: calculate the semantic relevance of the feature values ​​of each dimension of the user behavior feature vector with the ceramic culture attribute of the cross-modal fusion feature vector of the product, and use this relevance as the dimension optimization coefficient to correct the feature values ​​of each dimension of the user behavior feature vector dimension by dimension. By calculating the semantic similarity of ceramic cultural attributes, when correcting the coordinate distribution of feature vectors in the semantic space, the semantic similarity between the cross-modal fusion feature vectors of the product end and the user end and the features of each anchor point in the ceramic cultural attribute feature anchor point set is calculated. Taking the semantic weight of each cultural attribute anchor point in the marketing of ceramic cultural products as the correction benchmark, the coordinate values ​​of each feature vector in the marketing-specific feature semantic space are shifted and corrected towards the coordinate direction of the cultural attribute anchor point with high semantic similarity. The correction process keeps the dimension and modulus of the feature vector unchanged. The generation module is used to generate a set of personalized user demand tags for the corresponding ceramic culture product consumption dimension based on the encoded and mapped user interaction behavior feature data. The generation module generates a set of personalized user demand tags for ceramic culture products based on the encoded and mapped user feature data. Select the consumption dimension of ceramic culture products, then calculate the matching degree between user feature data and the demand of each consumption dimension, and finally take the consumption dimension with the demand matching degree higher than the preset threshold as the core demand dimension to generate the corresponding personalized demand tag set. The consumption dimensions of ceramic cultural products include cultural collection, craftsmanship appreciation, daily use, gift-giving, and intangible cultural heritage collection. When selecting a consumption dimension for ceramic cultural products, any one of these dimensions should be considered, and the degree of demand matching is required. , This represents the normalized user cross-modal fusion feature vector after encoding mapping. Let represent the standardized feature basis vector of the k-th consumption dimension of ceramic cultural products. This indicates the total number of cross-modal types, corresponding to the three modalities of visual, textual, and cultural semantics. Let represent the intermodal correlation coefficient between the m-th cross-modal feature and the k-th consumption dimension. This represents the number of dimensions in the unified feature space. This represents the feature space projection coefficient of the s-th dimension feature in the k-th consumption dimension; The above formula uses the inner product of the normalized cross-modal fusion feature vector of the user after encoding and mapping and the standardized feature basis vector of the corresponding consumption dimension as the basic matching basis. It also adds the product of the intermodal correlation coefficients of each cross-modal feature and the consumption dimension for gain adjustment, and makes corrections by combining the summation results of the feature space projection coefficients. The formula comprehensively calculates the matching degree between the user's needs and each consumption dimension, fully combining the characteristics of the cross-modal features of ceramic cultural products and the spatial projection law of the consumption dimension, and quantifies the fit between user needs and each consumption dimension in multiple dimensions. in, Based on the cross-modal feature dataset of ceramic cultural products of type m and the feature dataset of the consumption dimension of type k, the semantic correlation of ceramic culture is analyzed. The product of the feature co-occurrence probability of the two datasets and the semantic similarity of ceramic culture is calculated, and then the result is obtained after normalization of the ceramic culture attribute dimension. The s-th dimension feature vector in the cross-modal fusion feature space of ceramic cultural products after unification is orthogonally projected onto the standardized feature basis vector space of the k-th consumption dimension. The ratio of the magnitude of the projected vector to the magnitude of the standardized feature basis vector of the consumption dimension is then calculated. The module is used to build a set of product attribute tags that are adapted to the user's personalized needs tag set based on the coded and mapped ceramic culture product feature data. The construction module builds a product attribute tag set based on the encoded and mapped ceramic culture product feature data, adapting it to the user's personalized needs tag set. Based on the cross-modal fusion feature data of the product after encoding and mapping, the features are clustered and dimensionalized according to the exclusive attribute dimensions of ceramic culture products to obtain multi-dimensional attribute feature clusters of the product. Then, according to the core requirement dimensions of the user personalized requirement tag set, the multi-dimensional attribute feature clusters of the product are adapted and filtered to remove attribute features that have no semantic relationship with the core requirement dimensions. Finally, the filtered attribute features are tagged to generate product attribute tag set. The unique attributes of ceramic cultural products include kiln site attributes, craftsmanship attributes, vessel type attributes, cultural connotation attributes, material grade attributes, and intangible cultural heritage certification attributes. The matching module is used to generate a targeted recommendation list of ceramic culture products for each user based on the real-time matching relationship between the user's personalized demand tag set and the product attribute tag set; When the matching module generates a targeted recommendation list of ceramic culture products for a single user, it performs the generation operation based on a hierarchical matching strategy that prioritizes ceramic culture attributes. The first level matches the core cultural attributes of ceramic culture and intangible cultural heritage certification; the second level matches the technological attributes of kiln and craftsmanship; and the third level matches the practical attributes of vessel shape and material. The matching module first calculates the semantic matching degree of the tags under each level, and then obtains the comprehensive matching degree between the product and the user based on the weight of the levels. At the same time, it triggers matching updates based on real-time changes in user interaction behavior. Finally, it sorts the comprehensive matching degree from high to low and generates a targeted recommendation list of ceramic culture products for each user. Among them, the hierarchical weights are based on user-defined presets on the system side and are subject to the rule that the sum of the hierarchical weights is 1. The initial values ​​of the three hierarchical weights are equal and are adjusted in real time based on the semantic matching degree. That is, the semantic matching degree of each level is proportional to the hierarchical weight. The push module is used to push the generated targeted recommendation list to the corresponding user's online sales platform interface; When the push module pushes the targeted recommendation list to the user interface, it first calculates the push adaptation value for each ceramic culture product in the recommendation list, and then determines the display position, display section and display frequency of the product on the interface based on the push adaptation value. After the push module completes the push operation, it refreshes the system operation at specified time intervals. ; In the formula: The push adaptation value for the i-th ceramic culture product in the targeted recommendation list; Let be the feature vector representing the probability of reaching the target user for the i-th ceramic cultural product; Let be the feature vector of the fit between the i-th ceramic culture product and the user demand tag; Let i represent the cultural dissemination degree of the i-th ceramic cultural product; The cross-scenario push adaptation coefficient for the i-th ceramic culture product; This is the element-wise dot product operation of the eigenvectors; The above formula combines the dot product of product reach probability and demand fit, cultural dissemination and cross-scenario push adaptation coefficient to calculate the push adaptation value. This is different from the existing conventional method of determining push priority based solely on matching degree. It further aligns with the dual attributes of ceramic cultural products, making push planning more suitable for actual marketing scenarios. in, Based on statistical data of target users' historical clicks, browsing, and dwell times on similar ceramic cultural products across different interactive interfaces on e-commerce platforms, the reach probability value for each interface is obtained after normalization processing to determine... The values ​​of each dimension component; The semantic similarity of the attribute tag set of the i-th ceramic culture product with the core dimensions of the target user's personalized demand tag set is calculated one by one, and the semantic similarity value of each dimension is used as... The values ​​of the corresponding dimension components; The value range is the range of positive real numbers. The higher the degree of match between the display format of ceramic-specific attributes of ceramic cultural products and the content presentation requirements of different interactive scenarios on online sales platforms, as well as the user's browsing habits of ceramic cultural products in the corresponding scenarios, and the more complete and efficient the ceramic cultural semantic information of the product can be conveyed in that scenario, the better. The larger the value, the smaller the value; The acquisition module is interconnected with the encoding and mapping module via a wireless network. The encoding and mapping module is interconnected with the generation and construction modules via a wireless network. The construction module is interconnected with the matching module via a wireless network. The matching module is interconnected with the push module via a wireless network. The push module is interconnected with the acquisition module via a wireless network.

[0022] In this embodiment, the acquisition module collects and filters cross-modal representation data of ceramic cultural products from the online sales platform, as well as the full interactive behavior data of the corresponding users within the platform. The encoding and mapping module simultaneously performs unified-dimensional cross-modal feature encoding and semantic mapping processing on the collected and filtered cross-modal data of ceramic cultural products and user interactive behavior data. The generation module further generates a set of personalized user demand tags corresponding to the consumption dimension of ceramic cultural products based on the encoded and mapped user interactive behavior feature data. Then, the construction module constructs a set of product attribute tags that matches the set of personalized user demand tags based on the encoded and mapped ceramic cultural product feature data. The matching module generates a targeted recommendation list of ceramic cultural products for a single user based on the real-time matching relationship between the set of personalized user demand tags and the set of product attribute tags. Finally, the push module pushes the generated targeted recommendation list to the corresponding user's online sales platform interaction interface.

[0023] In the above embodiments, the system accurately mines users' personalized needs for ceramic culture products and adapts them to product attributes. Combined with a hierarchical matching strategy that prioritizes ceramic culture attributes, it achieves targeted product recommendations. It can also update matching results in real time based on user behavior and optimize product display according to the push adaptation value, making online recommendations for ceramic culture products more in line with user needs and improving marketing accuracy and user experience.

[0024] Regarding the system in the above embodiments, the following is an application example of the system: The XX online e-commerce platform brings together intangible cultural heritage ceramic merchants from well-known kilns such as Jingdezhen, Longquan Kiln, and Ru Kiln. The products cover multiple categories, including cultural collections, craft appreciation, daily use, and gifts. To improve the efficiency of precise marketing of ceramic cultural products and enhance the dissemination of ceramic culture, the platform has deployed this system. The following is a practical application case of the system for Mr. Li, a user of the platform.

[0025] The system's data acquisition module took the lead in collecting data. On the one hand, it collected cross-modal representation data of ceramic cultural products on the platform, specifically acquiring visual modal data (including high-definition image features of vessel outlines, decorative details, and glaze layers) of products such as Jingdezhen blue and white lotus pattern appreciation vases, Longquan celadon carved bowls, and Ru celadon sky-blue glazed teacups; textual modal data (including textual descriptions of kiln origins, techniques such as hand-throwing / blue and white painting / carving, and cultural connotations such as auspiciousness / refinement and leisure); and cultural semantic modal data (including information on intangible cultural heritage schools such as Jingdezhen blue and white firing techniques and Longquan celadon firing techniques, and semantic features of historical and cultural backgrounds such as Ming and Qing dynasty official kilns and Song dynasty Ru kilns). On the other hand, it collected full interactive behavior data of user Mr. Li over the past 30 days, including the time sequence data of his browsing 21 ceramic products, collecting 4 Jingdezhen blue and white intangible cultural heritage ceramics, adding 1 handmade blue and white tea set to his cart, and posting 2 comments related to ceramic craftsmanship, as well as detailed data such as the duration and frequency of each behavior. After data collection, the collection module performs data filtering. First, based on the core feature anchor points of ceramic cultural attributes constructed from ceramic intangible cultural heritage schools, kilns, historical and cultural backgrounds, and cultural connotations, cross-modal semantic similarity algorithms are used to calculate and remove data on five ordinary ceramic daily necessities that have no association with ceramic cultural attributes. Then, Mr. Li's behavioral data is standardized and quantified from three dimensions: duration, frequency, and depth of interaction behavior. The calculated comprehensive interaction behavior validity is 0.91, which is higher than the platform's preset threshold of 0.6. Therefore, all his interaction behavior data is retained, completing the data collection and filtering.

[0026] The system encoding and mapping module performs cross-modal feature encoding and semantic mapping on the filtered product and user data. In the feature encoding stage, multi-layer convolutional feature extraction and dimensionality reduction are performed on the product visual modality data to obtain a visual modality feature vector; word segmentation and cultural semantic word embedding encoding are performed on the product text modality data to obtain a text modality feature vector; cultural semantic entity linking and feature quantization encoding are performed on the product cultural semantic modality data to obtain a cultural semantic modality feature vector; and time-series feature serialization encoding and behavior validity weight quantization are performed on Mr. Li's behavioral time-series data, calculating his collection behavior validity weights as 0.35, add to cart as 0.29, browse as 0.16, comment as 0.12, share as 0.08, and place an order as 0 (no order behavior), forming a complete set of behavior validity weights and obtaining the user behavior modality feature vector. After entering the unified-dimensional feature transformation and fusion stage, the system projects the feature vectors of each modality onto the pre-constructed unified-dimensional feature space for ceramic culture product marketing, obtaining the same-dimensional feature vectors of each modality. Then, through a cross-modal attention mechanism, the contribution attention weights of each modality feature are calculated, resulting in 0.36 for visual modality, 0.33 for text modality, and 0.31 for cultural semantic modality. Based on these weights, the same-dimensional feature vectors of product categories are fused to generate a unified-dimensional cross-modal fused feature vector for the product side. Simultaneously, the semantic relevance of Mr. Li's user behavior feature vector is modified dimension by dimension in combination with the semantic relevance of product category features to generate a unified-dimensional cross-modal fused feature vector for the user side. Finally, semantic mapping is performed. The system uses the set of ceramic culture attribute feature anchor points, with intangible cultural heritage certification, kiln school, craftsmanship, and cultural connotation as the core, as the semantic reference benchmark. The fused feature vectors of the product side and the user side are synchronously mapped to the semantic space of ceramic culture product marketing features. The coordinate distribution of the feature vectors is corrected by calculating the semantic similarity of ceramic culture attributes, completing the semantic alignment of product features and user features in the same semantic space.

[0027] Based on Mr. Li's user feature data after encoding and mapping, the system generation module selects five consumption dimensions of ceramic cultural products: cultural collection, craft appreciation, daily use, gift giving, and intangible cultural heritage collection. It calculates the matching degree between the user feature data and the demand of each consumption dimension, and obtains the following results: cultural collection dimension 0.93, intangible cultural heritage collection dimension 0.90, craft appreciation dimension 0.78, gift giving dimension 0.45, and daily use dimension 0.39. Among them, the matching degree of the first three is higher than the platform's preset threshold of 0.7. Based on this, a personalized demand tag set is generated for Mr. Li, specifically: Jingdezhen kiln intangible cultural heritage ceramics, professional appreciation of blue and white porcelain craftsmanship, and high-end cultural collection ceramics.

[0028] The system construction module is based on the cross-modal fusion feature data of ceramic cultural products after encoding and mapping. It performs feature clustering and dimensionality subdivision according to six exclusive attribute dimensions of ceramic cultural products: kiln site attributes, craft techniques attributes, vessel type attributes, cultural connotation attributes, material grade attributes, and intangible cultural heritage certification attributes, to obtain multi-dimensional attribute feature clusters for each product. Then, based on the core demand dimensions of Mr. Li's personalized demand tag set, it filters out product attribute features that are not semantically related to Jingdezhen kiln sites, blue and white porcelain techniques, and cultural collections, such as features related to Longquan celadon, Ru kiln sky blue glaze, and ordinary daily-use ceramics. The filtered product attribute features are then tagged to construct a product attribute tag set that is highly adapted to Mr. Li's needs, specifically including: Jingdezhen kiln site, intangible cultural heritage blue and white porcelain firing techniques, hand-thrown + blue and white porcelain painting techniques, appreciation vase / ornament vessel type, high white clay raw material, auspicious cultural connotations, and cultural collection-grade ceramics.

[0029] The system matching module generates a targeted recommendation list based on Mr. Li's personalized demand tag set and product attribute tag set, using a hierarchical matching strategy that prioritizes ceramic cultural attributes. This strategy is divided into three levels: the first level is the matching of core cultural attributes such as cultural connotation and intangible cultural heritage certification; the second level is the matching of technological attributes such as kiln and craftsmanship; and the third level is the matching of practical attributes such as vessel shape and material. The initial weight of each level is 0.33. The system first calculates the semantic matching degree of tags at each level, obtaining a matching degree of 0.96 for core cultural attributes, 0.93 for craftsmanship attributes, and 0.88 for practical attributes. Then, it adjusts the level weights in real time based on the semantic matching degree of each level. After adjustment, the weights for core cultural attributes are 0.35, craftsmanship attributes are 0.34, and practical attributes are 0.31. Based on the adjusted weights, the system calculates the overall matching degree between each product and Mr. Li. Specifically, the Jingdezhen blue-and-white lotus-patterned intangible cultural heritage appreciation vase has a matching degree of 0.95, the Jingdezhen handmade blue-and-white landscape-patterned collectible vase has a matching degree of 0.92, and the Jingdezhen blue-and-white lotus-patterned handmade tea set has a matching degree of 0.88. The overall matching degree of the remaining products is below 0.85. The system sorts the overall matching degrees from high to low, generating a personalized ceramic culture product recommendation list for Mr. Li. Simultaneously, the system monitors Mr. Li's interactive behavior in real time; if he makes new collections, adds to cart, or comments, the recommendation list will be immediately updated.

[0030] The system's push module calculates the push compatibility value for products in the generated targeted recommendation list. The compatibility values ​​are: Jingdezhen blue-and-white lotus-patterned intangible cultural heritage appreciation vase (0.94), Jingdezhen handmade blue-and-white landscape-patterned collectible vase (0.90), and Jingdezhen blue-and-white lotus-patterned handmade tea set (0.85). Based on these compatibility values, the system determines the display position, section, and frequency of each product on Mr. Li's platform interface: the Jingdezhen blue-and-white lotus-patterned intangible cultural heritage appreciation vase, with the highest compatibility value, is displayed first in the "Selected Intangible Cultural Heritage Ceramics" core section on the platform's homepage, appearing prioritized every hourly refresh; the Jingdezhen handmade blue-and-white landscape-patterned collectible vase is displayed second in this section, with the next highest frequency; and the Jingdezhen blue-and-white lotus-patterned handmade tea set is displayed first in the platform's "Blue-and-White Craft Ceramics" exclusive section, refreshed every 30 minutes. After the push operation is completed, the system refreshes the overall operating status at specified 30-minute intervals, continuously collecting Mr. Li's new interactive behavior data and the latest data on ceramic cultural products on the platform, achieving dynamic optimization of Mr. Li's ceramic cultural product recommendations.

[0031] In summary, the system described in the above embodiments, through targeted collection and two-layer screening of multimodal representation data of ceramic cultural products and user interaction behavior data, can accurately eliminate product data without ceramic cultural attributes and user data without effective interaction, thus solidifying the accuracy and effectiveness of data application. Simultaneously, by independently encoding multimodal features and fusing them with unified-dimensional feature transformation, combined with a cross-modal attention mechanism, it assigns appropriate marketing contribution weights to each modality feature, ensuring precise semantic alignment between product and user features in the exclusive semantic space of ceramic cultural product marketing. This allows feature representations to align with the exclusive attributes and consumer demand dimensions of ceramic cultural products. Furthermore, based on feature data, it accurately mines users' personalized core needs in ceramic cultural product consumption, constructs highly compatible product attribute tags, and generates a recommendation list using a hierarchical matching strategy prioritizing cultural attributes and incorporating real-time user behavior changes. The hierarchical weights are dynamically adjusted according to semantic matching degree to maintain the alignment between product recommendations and user needs. In addition, by calculating push adaptation values, it optimizes the product's display position, section, and frequency on the interactive interface, making product reach more aligned with user browsing habits and effectively improving the online marketing conversion efficiency of ceramic cultural products.

[0032] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-modal, large-scale model-driven marketing system for ceramic cultural products, characterized in that: include: The data collection module is used to collect and filter cross-modal representation data of ceramic cultural products on online sales platforms, as well as the full interactive behavior data of corresponding users within the platform. The encoding and mapping module is used to perform unified-dimensional cross-modal feature encoding and semantic mapping processing on the collected and filtered cross-modal data of ceramic cultural products and user interaction behavior data. The generation module is used to generate a set of personalized user demand tags for the corresponding ceramic culture product consumption dimension based on the encoded and mapped user interaction behavior feature data. The module is used to build a set of product attribute tags that are adapted to the user's personalized needs tag set based on the coded and mapped ceramic culture product feature data. The matching module is used to generate a targeted recommendation list of ceramic culture products for each user based on the real-time matching relationship between the user's personalized demand tag set and the product attribute tag set; The push module is used to push the generated targeted recommendation list to the corresponding user's online sales platform interface.

2. The cross-modal large-model driven ceramic culture product marketing system according to claim 1, characterized in that, The cross-modal representation data of ceramic cultural products collected by the acquisition module includes visual modal data, text modal data, and cultural semantic modal data specific to ceramic cultural products. The visual modal data includes image feature data of ceramic shapes, patterns, and glazes; the text modal data includes textual descriptions of ceramic kilns, techniques, and cultural connotations; and the cultural semantic modal data includes semantic feature data of the intangible cultural heritage school to which the ceramics belong and their historical and cultural background. The full user interaction behavior data includes time-series data of user behavior such as browsing, collecting, adding to cart, placing orders, commenting, and sharing ceramic culture products. The filtering operation of the collection module is to perform an initial screening of cross-modal representation data based on the relevance of ceramic culture attributes, and a secondary screening of user interaction behavior data based on the validity of behavior, removing product data that is not associated with ceramic culture attributes and user data that has no valid interaction behavior.

3. The cross-modal large-model driven ceramic culture product marketing system according to claim 1, characterized in that, The encoding and mapping module performs cross-modal feature encoding and semantic mapping processing, including the following stages: Feature Encoding: Feature extraction and encoding were performed on the visual modal data, textual modal data, cultural semantic modal data, and time-series data of all user interaction behaviors of ceramic cultural products. Multi-layer convolutional feature extraction and dimensionality reduction were performed on the image feature data of ceramic shapes, patterns, and glazes to obtain visual modal feature vectors. The textual descriptions of ceramic kilns, techniques, and cultural connotations are segmented and encoded using cultural semantic terms to obtain text modal feature vectors. The semantic feature data of ceramic intangible cultural heritage schools and historical and cultural backgrounds are linked with cultural semantic entities and quantified and encoded to obtain cultural semantic modal feature vectors. The user behavior modal feature vector is obtained by performing temporal feature serialization encoding and behavior validity weight quantization on the time-series data of user browsing, favorites, adding to cart, placing orders, commenting, and sharing behaviors. ; Unified Dimension Feature Transformation and Fusion: A unified dimension feature space for ceramic culture product marketing is pre-constructed, and modal feature projection matrices for visual, textual, cultural semantics, and user behavior are configured respectively. , , , By projecting the corresponding projection matrix onto this unified-dimensional feature space, we obtain the same-dimensional feature vectors for each modality. Then, the attention weights contributing to the marketing of ceramic cultural products are calculated using a cross-modal attention mechanism. Based on these weights, the attention weights of product category feature vectors of the same dimension are then applied. The features are fused to generate a unified cross-modal fusion feature vector for the product. For user behavior feature vectors of the same dimension By optimizing the semantic relevance of product category features, a unified cross-modal fusion feature vector for the user end is generated. ; Semantic mapping: Constructing a set of anchor points for ceramic cultural attributes, which includes at least four core ceramic cultural attributes: intangible cultural heritage certification, kiln origin and school, craftsmanship, and cultural connotation. Using this anchor point set as a semantic reference benchmark, cross-modal fusion feature vectors from the product side are then used. Cross-modal fusion feature vectors with the user end Synchronously map to the preset semantic space of ceramic culture product marketing features, and then correct the coordinate distribution of feature vectors in the semantic space through the semantic similarity calculation of ceramic culture attributes, so as to complete the semantic alignment of product features and user features in the same semantic space.

4. The cross-modal large-model driven ceramic culture product marketing system according to claim 3, characterized in that, When calculating the attention weights of each modal feature's contribution to the marketing of ceramic cultural products using the cross-modal attention mechanism, the following rules apply: ； In the formula: The attention weights are contributed to the modal feature vector of the m-th product class in the same dimension. Let be the correlation degree of the ceramic culture attribute of the m-th modal feature; Let be the marketing feature expressibility of the m-th modality.

5. A cross-modal large-model driven marketing system for ceramic cultural products according to claim 3, characterized in that, The same-dimensional user behavior feature vector The optimization operation based on the semantic relevance of product category features is as follows: calculate the semantic relevance of the feature values ​​of each dimension of the user behavior feature vector with the ceramic culture attribute of the cross-modal fusion feature vector of the product, and use this relevance as the dimension optimization coefficient to correct the feature values ​​of each dimension of the user behavior feature vector dimension by dimension. By calculating the semantic similarity of ceramic cultural attributes, when correcting the coordinate distribution of feature vectors in the semantic space, the semantic similarity between the cross-modal fusion feature vectors of the product end and the user end and the features of each anchor point in the ceramic cultural attribute feature anchor point set is calculated. Taking the semantic weight of each cultural attribute anchor point in the marketing of ceramic cultural products as the correction benchmark, the coordinate values ​​of each feature vector in the marketing-specific feature semantic space are shifted and corrected towards the coordinate direction of the cultural attribute anchor point with high semantic similarity. The correction process keeps the dimension and modulus of the feature vector unchanged.

6. The cross-modal large-model driven marketing system for ceramic cultural products according to claim 1, characterized in that, The generation module generates a set of personalized user demand tags for ceramic cultural products based on the encoded and mapped user feature data. Select the consumption dimension of ceramic culture products, then calculate the matching degree between user feature data and the demand of each consumption dimension, and finally take the consumption dimension with the demand matching degree higher than the preset threshold as the core demand dimension to generate the corresponding personalized demand tag set. The consumption dimensions of ceramic cultural products include cultural collection, craftsmanship appreciation, daily use, gift-giving, and intangible cultural heritage collection. The degree of demand matching... , This represents the normalized user cross-modal fusion feature vector after encoding mapping. Let represent the standardized feature basis vector of the k-th consumption dimension of ceramic cultural products. This indicates the total number of cross-modal types, corresponding to the three modalities of visual, textual, and cultural semantics. Let represent the intermodal correlation coefficient between the m-th cross-modal feature and the k-th consumption dimension. This represents the number of dimensions in the unified feature space. This represents the feature space projection coefficient of the s-th dimension feature in the k-th consumption dimension.

7. The cross-modal large-model driven ceramic culture product marketing system according to claim 1, characterized in that, The construction module constructs a product attribute tag set that matches the user's personalized needs tag set based on the encoded and mapped ceramic culture product feature data. Based on the cross-modal fusion feature data of the product after encoding and mapping, the features are clustered and dimensionalized according to the exclusive attribute dimensions of ceramic culture products to obtain multi-dimensional attribute feature clusters of the product. Then, according to the core requirement dimensions of the user personalized requirement tag set, the multi-dimensional attribute feature clusters of the product are adapted and filtered to remove attribute features that have no semantic relationship with the core requirement dimensions. Finally, the filtered attribute features are tagged to generate product attribute tag set. The exclusive attribute dimensions of the ceramic cultural products include kiln site attributes, craftsmanship attributes, vessel type attributes, cultural connotation attributes, material grade attributes, and intangible cultural heritage certification attributes.

8. The cross-modal large-model driven ceramic cultural product marketing system according to claim 1, characterized in that, When the matching module generates a targeted recommendation list of ceramic culture products for a single user, it performs the generation operation based on a hierarchical matching strategy that prioritizes ceramic culture attributes: The first level matches the core cultural attributes of ceramic culture and intangible cultural heritage certification; the second level matches the technological attributes of kiln and craftsmanship; and the third level matches the practical attributes of vessel shape and material. The matching module first calculates the semantic matching degree of the tags under each level, and then obtains the comprehensive matching degree between the product and the user based on the weight of the levels. At the same time, it triggers matching updates based on real-time changes in user interaction behavior. Finally, it sorts the comprehensive matching degree from high to low and generates a targeted recommendation list of ceramic culture products for each user.

9. A cross-modal large-model driven marketing system for ceramic cultural products according to claim 1, characterized in that, When the push module pushes the targeted recommendation list to the user interface, it first calculates the push adaptation value for each ceramic culture product in the recommendation list, and then determines the display position, display section and display frequency of the product on the interface based on the push adaptation value. After the push module completes the push operation, it refreshes the system operation at a specified time interval. ； In the formula: The push adaptation value for the i-th ceramic culture product in the targeted recommendation list; Let be the feature vector representing the probability of reaching the target user for the i-th ceramic cultural product; Let be the feature vector of the fit between the i-th ceramic culture product and the user demand tag; Let i represent the cultural dissemination degree of the i-th ceramic cultural product; The cross-scenario push adaptation coefficient for the i-th ceramic culture product; This is the element-wise dot product operation of the eigenvectors.

10. A cross-modal large-model driven marketing system for ceramic cultural products according to claim 1, characterized in that, The acquisition module is interconnected with an encoding and mapping module via a wireless network. The encoding and mapping module is interconnected with a generation module and a construction module via a wireless network. The construction module is interconnected with a matching module via a wireless network. The matching module is interconnected with a push module via a wireless network. The push module is interconnected with the acquisition module via a wireless network.

Images